Comparative Analysis of Feature Extraction Methods on Convolutional Neural Network Models for Classification of Healthy and Unhealthy Cacao Pods

Abstract

Objective classification of cacao pods can support consistent assessment of pod health under field conditions. This study developed and compared convolutional neural network (CNN) models for classifying healthy and unhealthy cacao pods using raw images and three feature extraction approaches: Local Binary Pattern (LBP), Edge Feature (EF), and Gabor Filter (GF). A total of 500 cacao pod images were collected using a high-resolution DSLR camera and labelled by expert cacao farmers, comprising 250 healthy and 250 unhealthy pods. Images were preprocessed, and the datasets were partitioned into 70% training, 20% validation, and 10% test sets. Four CNN configurations were evaluated: a baseline CNN using raw images and CNN models using LBP-, EF-, and GF-processed inputs. The models were trained for 100 epochs and evaluated using confusion-matrix-derived classification metrics. The CNN with Edge Feature preprocessing achieved the highest reported performance, with 96.15% accuracy, a Kappa score of 0.923, and an F-measure of 96.15%. The LBP and Gabor Filter configurations achieved accuracies of 90.38% and 88.46%, respectively, while the baseline CNN achieved 69.23%. The findings indicate that feature extraction can improve CNN-based cacao pod classification, with edge-based preprocessing providing the strongest performance among the configurations evaluated. Further validation across broader field conditions is warranted before deployment.